[cristiangirlea]

Fixes #104.

Three defects found by review after the logging and owner-cap merges, none of which leaks anything:

• A reply that was never delivered could be logged as sent: the status was recorded before the write and the request line written from a finally, so a client reset mid-reply still logged a 200, and the busy 503 was logged even when its non-blocking send raised. A reply is now recorded as sent only after its body is written (the send_error override does the same after the standard library's whole reply, including a headers-only HEAD refusal), and the busy 503 is logged only when the send delivered every byte; otherwise no request line, as for a connection that ends before any reply. The protocol guide's Logging section says so.
• The drain line could be lost after a second signal, because server_close waited one second for the drain thread. The thread now publishes its fields and sets an Event before calling shutdown(), and server_close writes the line from that result if the thread has not, exactly once and before stop, waiting up to five seconds; past that, stop proceeds without the line, which the guide also says.
• With BRIDGE_LOG=off the request timer, the route and method helpers, the job timings and the drain's snapshot still ran; a cheap log.enabled() check now skips them.

Tests, seen failing first where the behaviour was wrong: a reply whose first write fails writes no request line, nor does a failed or short busy send; a drain thread held past one second still writes exactly one drain line with interrupted: true before stop, and past the five-second wait writes none; with logging off the timer and field construction are skipped, not just the output; captured lines for a 429 and a 503 draining with status, owner, schema and no sentinel; and the owner cap with cancelled jobs (concurrency 1, cap 2: cancelled jobs keep the cap until one expires, while another owner still submits).

Files: worker/ai_bridge/server.py, worker/ai_bridge/jobs.py, worker/ai_bridge/log.py, tests/python/test_server.py, tests/python/test_jobs.py, tests/python/test_log.py, docs/protocol.md, docs/testing.md, CHANGELOG.md.

Checks: the compose suites could not run locally because Docker Desktop stopped responding on the development machine mid-run ("Docker Desktop is unable to start"); 51 targeted tests ran green in the containers before that. The whole Python suite was then run directly on the Windows host with Python 3.13: 415 tests, with the same fifteen platform-only failures as main on that host (file-mode and shell-script cases that the containers cover) and one order-dependent error in test_a_busy_503_is_logged_without_anything_read, which passes when run alone. The CI gate on this pull request is therefore the check of record for this head; it runs the full sequence including ruff and mypy at strict. Review: independent review by Codex (gpt-6.1-sol): no defect, every listed test present and distinguishing main's behaviour.